SOCK enables training generative models for financial time series by making random convolutional features differentiable—a key innovation that lets you supervise generators without the information loss of signature-based methods.
This paper solves the problem of generating realistic financial time series from limited historical data by introducing SOCK, a differentiable random convolutional feature map. Instead of using path signatures (which lose information at practical depths), SOCK lets generators learn by matching feature representations of real and generated data, avoiding overfitting and discriminator memorization.